The features are directions in the activation space.
Based on the discussion in the cohort, it seemed like, on a particular layer, directions are the possible vectors with neuron as the basis.
But in that scenario, the polysemanaticity should be the norm and not a surprising fact.
Sept 4
Sociolinguistics foundations of AI
No insights into the methods that can be employed to access those biases
Interesting ideas but not developed further. to explore how they can be useful
No claims are proven, and since those conclusions are already part of the general knowledge among practitioners of LLMs
Real contribution being only to provide a taxonomy, which is useful and can be further developed but the current article falls short on providing anything useful or a pathway to how it can become useful.
Sept 3
Sociolinguistics foundations of AI
Considers LLMs as modelling variety of language
Provides a taxonomy to analyze the bias and other aspects of language generated by LLMs
